Responses API ignores replayed `mcp_call` input items unless they reference a stored response

Summary

With store: false, the docs say to
replay a response’s output items as the next request’s input. For hosted MCP that doesn’t work: an mcp_call input
item is dropped entirely, including its arguments, output and error, unless its id names a stored response. The
model never sees an earlier MCP result, so in a multi-step agent loop it invents values from that result or calls the
tool again.

Repro (gpt-5.4, reasoning effort low, public DeepWiki MCP server, 3 trials each)

  1. Request with tools: [{"type":"mcp","server_label":"deepwiki","server_url":"https://mcp.deepwiki.com/mcp","require_approval":"never"}]
    and a forced ask_wiki_question call. Keep the output items.
  2. Follow-up: the original user message, the output items, then
    “Quote the first twelve words of the answer deepwiki returned earlier. If you can’t see it, reply UNKNOWN.”
Follow-up Model saw the result input_tokens
store: true, previous_response_id 3/3 738
store: true, items resent with their stored ids 3/3 714
store: false, items resent (reasoning id kept, others stripped) 0/3 243
store: false, items resent with the mcp_call removed 0/3 243
store: false, items resent with their ids 0/3 —
store: true, mcp_call without an id 0/3 —
store: false, the same text as function_call + output 3/3 —

An inline mcp_call costs exactly as many input tokens as no item at all, so it is dropped before the prompt is
built. A failed call’s mcp_tool_execution_error content and the call’s arguments are dropped too. An inline
mcp_list_tools item, by contrast, is honoured: the tools aren’t listed again.

After DELETE /v1/responses/{id}, resending the stored ids returns 400 item_conflict, so storing responses only
moves the problem to day 31. Zero Data Retention organisations can’t store responses at all.

Expected

An mcp_call input item with output (or error) is shown to the model like a function_call + function_call_output
pair, whether or not it references a stored response.

Workaround in use

Each finished mcp_call is followed by an assistant message (phase: "commentary") that restates the call and its
output.

The reasoning_id is not a server-side retained item that you can recall. You must send back the encrypted reasoning.

There’s basically one way to make a follow-up call correctly.

What I do not see above:

  • use the “include” parameter, with the array of strings including " reasoning.encrypted_content" for assurance against a changing API;
  • replay the entire output of a response back as additional input items, including reasoning items with “encrypted_content” preserved, include compaction items. Basically, do not alter any of the output as you seem to be keen to doing, simply replay it in full.
  • Besides effort, employ "reasoning":{"context": "all_turns"} in calls so reasoning is not dropped (and backport to gpt-5.4 and see return);
  • phase included in the output will also be replayed (include support if not a direct echo).
  • include every output, including every mcp event type such as approvals.
  • don’t switch between models in a self-managed chat context. This will drop all encrypted content. Run the whole sequence again if testing.

Report on: I would check the switch between gpt-5.4 and gpt-5.6, with its default of all_turns of reasoning. That model also may promote more lossless full input even with the full unaltered replay. Then you can report on model specificity that may cause an issue. MCP tools support (or any tools at all) is now reported on per-model documentation starting at gpt-5.4 but you can’t quite trust that dozens of per-model gates are being maintained correctly.